Explore the vulnerabilities of AI-powered game bots in this 40-minute Black Hat conference talk. Delve into the world of deep reinforcement learning (DRL) algorithms and their application in complex games like StarCraft. Learn about modeling and solving reinforcement learning problems, training DRL bots, and various attack methods including perturbation-based and adversarial agent attacks. Examine quantitative evaluations of winning rates, view demo examples, and discuss potential defense strategies against these exploits. Gain insights from experts Xinyu Xing, Wenbo Guo, Xian Wu, and Jimmy Su as they reveal the blind spots of AI in gaming and demonstrate how to exploit them.
Ruling StarCraft Game Spitefully - Exploiting the Blind Spot of AI-Powered Game Bots
Overview
Syllabus
Intro
DRL in Games
Modeling an RL Problem
Solving an RL Problem
DRL-powered Games
Training an DRL Bot
Perturbation-based Attacks
Adversarial Agent Attack
Our attack
Quantitively Evaluation • Comparison of winning rates.
Demo Examples
A Potential Defense
Taught by
Black Hat
